Reducing Cache Pollution via Dynamic Data Prefetch Filtering

摘要

In order to bridge the gap of the growing speed disparity between processors and their memory subsystems, aggressive prefetch mechanisms, either hardware-based or compiler-assisted, are employed to hide memory latencies. As the first-level cache gets smaller in deep submicron processor design for fast cache accesses, data cache pollution caused by overly aggressive prefetch mechanisms will become a major performance concern. Ineffective prefetches not only offset the benefits of benign prefetches due to pollution but also throttle bus bandwidth, leading to an overall performance degradation. In this paper, we propose and analyze a number of hardware-based prefetch pollution filtering mechanisms to differentiate good and bad prefetches dynamically based on history information. We designed three prefetch pollution filters organized as a one-level, two-level, or gshare style. In addition, we examine two table indexing schemes: per-address (PA) based and program counter (PC) based. Our prefetch pollution filters work in tandem with both hardware and software prefetchers. As our analysis shows, the cache pollution filters can reduce the ineffective prefetches by more than 90 percent and alleviate the excessive memory bandwidth induced by them. Also, the performance can be improved by up to 16 percent when our filtering mechanism is incorporated with aggressive prefetch filters as a result of reduced cache pollution and less competition for the limited number of cache ports. In addition, a number of sensitivity studies are performed to provide more understandings of the prefetch pollution filter design